Most long-read SV callers have parameters for the minimum read length that is considered when calling variants e.g 500bp or 1000bp. If you had a good N50 you might even consider raising this to 2000 or 5000bp, as I've seen done before.
But is there any strong disadvantage to lowering this number to 100 or 200bp, given these shorter reads may still be informative, particularly for small SVs and samples with low average read length?
Yes good idea. I guess the quality of the SV calls and as a proxy for quality the length and distribution of variants called might be informative. But without a truthset it might be difficult to definitively say how much better the calls are when excluding the shortest reads.